Local Cluster First Load Sharing Policy for Heterogeneous Clusters

This paper studies the load sharing problem among heterogeneous cluster systems. The heterogeneous clusters we consider are time-sharing, and the computers in these clusters have different CPU powers and memory capacities. Load sharing means even workloads among all coordinated computers in the system. As some nodes suffer from high loading, it is necessary to migrate some jobs to the nodes with light loading. Recently, many load sharing policies, such as CPU-based and Memory-based policies, have been proposed for single-cluster system. However, they cannot be directly applied to larger-scale heterogeneous cluster systems. Without careful assignment, the cluster systems will waste time in moving migrated jobs to remote cluster, and thus reduce the performance of the cluster systems. In this paper we propose a simple but effective load sharing policy, local cluster first (LCF) policy, to improve the performance of various load sharing algorithms on heterogeneous cluster systems. The idea of LCF is that when some nodes suffer from heavy CPU (or/and memory) loading, the migrating procedure is carried on hierarchically. Since the heavy loading computer moves most of the migrated jobs to computers in the local cluster, the remote cluster migrations are reduced. We use a trace-driven simulator to simulate a heterogeneous cluster system. Compared with other load sharing policies, LCF reduces a lot of remote cluster migrations. As a result, LCF policy may improve the performance of load sharing on heterogeneous cluster systems.